Figure 2
At the top left, the textbox “Understanding Online Product Return Behavior” is prominently displayed. It contains four sections, and each is enclosed in dashed rectangles. A downward arrow from this textbox leads to the first section. The first section, on the left side, is labeled “(1) Data collection and input variable construction,” which starts from a textbox “Amazon dot com” as the data source. This textbox is divided into two ovals labeled “57,305 unique products” and “209,761 reviews.” The combined result from these two leads to another textbox, “Data cleaning,” which is further divided into a textbox “Topic modelling” on the left, an oval in the center that is “Price,” and an oval on the right that is “Sales rank.” A downward arrow from topic modelling leads to another box, “Information retrieval,” followed by an oval, “Product quality attributes.” A rightward arrow from this leads to another box labeled “Data merging,” which is also connected by “Price” and “Sales rank.” A rightward arrow from “Data cleaning” leads to the second section, which is titled “(2) Output variable construction.” It starts from the oval “10 percent reviews,” followed by the textbox “Human and machine labeling and validation,” and “S L and S S L model training.” An arrow from this box leads to another textbox positioned on the top right of the second section, which is labeled “Unseen data validation.” A downward arrow from this leads to “Model deployment,” followed by an oval “90 percent reviews.” The first and second sections lead to a dashed rectangle of “Data integration,” followed by the text “Create a new dataset combining input variables and output variable.” A rightward arrow from the data integration leads to the third section, which is titled “(3) Model selection.” It contains two textboxes labeled “S L model fitting” and “10-fold cross validation.” The combined result from these leads to another textbox labeled “Model evaluation,” followed by another textbox labeled “Variable importance ranking and Permutation test,” positioned at the top. An upward arrow from the third section leads to the fourth section, which is labeled “(4) Non-Assertive O P R B exploration.” It contains four textboxes arranged vertically and connected by upward arrows. The labels of the textboxes from bottom to top are “2 D Partial dependence plot,” “Variable interaction strength,” “3 D Partial dependence plot,” and “Result interpretations and recommendations.” A leftward arrow from the fourth section is connected to the main textbox “Understanding Online Product Return Behavior.”Proposed methodological framework. Source: Authors’ own creation